Research Byte
Published in the RSAA Lunations
Vol1 Issue22 1–30 November 2021
Improving adaptive optics performance using new data-driven connectionist methods
The largest ground-based telescopes will soon reach ~40m diameter and provide the angular resolution and collecting area required to detect the first stars and first galaxies as well as faint rocky exoplanets through direct imaging. But to reach the required resolution and contrast, they must overcome optical distortions induced by atmospheric turbulence. In order to compensate for such distortions, Adaptive Optics (AO) technologies were developed for astronomy starting in the late 1980s, and are now essential for the current largest optical telescopes. In its simplest form, an AO system is composed of a wavefront sensor (WFS) used to measure atmospheric distortions at a high frame rate, which are then compensated with a deformable mirror (DM). The sub-system linking those components, responsible for interpreting wavefront measurements into actual commands to actuators of the DM, is the real-time controller (RTC). It must operate at high speed (~ kHz rate) to catch up with the rapidly changing optical turbulence.
Whether applied directly to AO control, or otherwise in post-processing for PSF reconstruction, the classical approaches to phase retrieval are based on linear models of the relationship between the sensor data and the phase of the light wave. In both cases, phase retrieval is based on careful modeling of the AO system error budget and is solved through a linear approach, sometimes regularized with a given prior on the turbulence statistics. For instance, in present-day conventional AO systems, the RTC follows a well-defined linear control scheme: input measurements vectors from sensors are multiplied by a control matrix to produce an output DM control command vector.
In order to beat the current limitations of existing model-based techniques, we are developing new data-driven connectionist architectures, based on modern Machine Learning techniques using Deep Neural Networks. These must be able to cope with varying operating conditions for the AO system and various data sources including WFS and science instruments as well as non-functional constraints such as real-time operations. Our challenge is to develop novel pragmatic techniques that address non-linearities in an integrated data-driven framework, for the calibration and control of the AO system as well as data post-processing tasks. Developing data-driven and constantly learning phase retrieval strategies at both AO control and post-processing levels is a totally disruptive approach as compared to currently implemented techniques, which are limited by both the calibration accuracy and the fidelity of our models to the actual physical processes. As of today, while massive amounts of data are available from telescopes currently in operation, these kinds of data heavy approaches have never been implemented at scale.
So far, this research has been following two paths. On the one hand, we have been exploring the use of reinforcement learning to build an agile and predictive AO controller. This solution is based on a multi-agent model-free strategy working as an additive corrector on top of a linear controller, with an autoencoder at the frontend to mitigate the impact of noise in WFS data. We have shown that this solution allows us to increase the AO performance over a classical integrator controller and has similar or better performance than state-of-the-art model-based predictive controllers, with the added benefit of being data-driven, trained online during operations, without requiring any prior on the turbulence statistical properties and corresponding parameters. Time-to-solution for the combination of the denoising autoencoder and the Multi-Agent Reinforcement Learning controller is below 2 ms considering perfectly parallelised architecture making it compatible with a realistic operational scenario. We have also assessed the performance under changing turbulence conditions for which this new controller appears to be robust enough. The next step is to implement this new approach on actual experiments, in the lab and possibly on-sky, in order to evaluate the performance and behaviour (e.g. training time, stability, changing atmospheric conditions) in a real setting and increase the readiness level of this technique for future extreme AO instruments.
On the other hand, we are investigating the use of conditional adversarial artificial neural network architectures to predict phase using the wavefront sensor data from a closed-loop AO system. So far, we have shown that such an approach can utilise high order information from raw WFS data that is inaccessible with current algorithms and theoretical models. With this high order reconstruction, our translational network can accurately estimate the wavefront from just the WFS image with no loss of accuracy and has potential to improve the long exposure PSF estimation process, and so increase the sensitivity in astronomical images after post-processing. Compared to the state-of-the-art model-based approach, our method is not explicitly limited by modelling assumptions and is conceptually simple and flexible. On key quality metrics, specifically the Strehl ratio and halo distribution, our approach achieves results as good as, and sometimes better than the model-based baseline. Importantly for future application to the real-time control of AO instruments, inference of phase from wavefront sensor data using our approach occurs in a fraction of a millisecond on commodity hardware. This could potentially have a strong impact on how people will design future AO systems, either as a cost saving strategy (using cheaper components) or as a science enabler, increasing sky coverage or building a more potent system providing extreme AO correction, with the same sky coverage as classical AO.
This work is mainly led by two PhD students (Jeffrey Smith and Bartolomeu Poumulet), in collaboration with ANU-CECS, the Barcelona Supercomputing Center (BSC) and the Polytechnic University of Catalonia (UPC). The team includes Charles Gretton and Felipe Trevisan (CECS), Eduardo Quinones (BSC), Mario Martin (UPC), Jesse Cranney and Damien Gratadour (RSAA).
Damien Gratadour
